SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning

In three linesSLAT is an RL framework that reduces redundancy in chain-of-thought reasoning by selectively suppressing low-utility segments. On standard benchmarks, the method achieves 50% reasoning length reduction while maintaining competitive accuracy.

Summary generated by Claude — human-verified